Abstract: The online transformer fault diagnosis can realize the real-time monitoring of the transformer status and adopt proper prior intervening, a great significance for power grid. This deduces that the intelligent transformer fault diagnosis is a key part of the smart grid construction. The fault features derived from the online monitoring of the oil dissolved gas have become the information source of the intelligent diagnosis method, and the quality of the gas directly affects diagnostic effect. At present, there are many types of transformer fault features, and the intelligent algorithms are often used to select the features. However, each single feature selection method has its own characteristics in both feature numbers and diagnostic effects. To combine these advantages, a new selection method of transformer features is proposed here. Based on the IEC TC10 database and the public literature samples, the characteristics of those single feature selection methods are analyzed and a new feature fusion optimization method is introduced. Through the field sample verification, the feature ranking result based on the fusion method is more reasonable than the individual ranking of each algorithm, and the selected feature subsets used by the new method has obvious advantages than the single method and the traditional gas ratio method.